Re: of_node_put() usage is buggy all over drivers/of/base.c?!

[Frank Rowand]

On 8/16/21 2:20 PM, Rob Herring wrote:
> On Mon, Aug 16, 2021 at 10:14 AM Frank Rowand <frowand.list@xxxxxxxxx> wrote:
>>
>> On 8/16/21 9:46 AM, Vladimir Oltean wrote:
>>> Hi Frank,
>>>
>>> On Mon, Aug 16, 2021 at 09:33:03AM -0500, Frank Rowand wrote:
>>>> Hi Vladimir,
>>>>
>>>> On 8/13/21 8:01 PM, Vladimir Oltean wrote:
>>>>> Hi,
>>>>>
>>>>> I was debugging an RCU stall which happened during the probing of a
>>>>> driver. Activating lock debugging, I see:
>>>>
>>>> I took a quick look at sja1105_mdiobus_register() in v5.14-rc1 and v5.14-rc6.
>>>>
>>>> Looking at the following stack trace, I did not see any calls to
>>>> of_find_compatible_node() in sja1105_mdiobus_register().  I am
>>>> guessing that maybe there is an inlined function that calls
>>>> of_find_compatible_node().  This would likely be either
>>>> sja1105_mdiobus_base_tx_register() or sja1105_mdioux_base_t1_register().
>>>
>>> Yes, it is sja1105_mdiobus_base_t1_register which is inlined.
>>>
>>>>>
>>>>> [  101.710694] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:938
>>>>> [  101.719119] in_atomic(): 1, irqs_disabled(): 128, non_block: 0, pid: 1534, name: sh
>>>>> [  101.726763] INFO: lockdep is turned off.
>>>>> [  101.730674] irq event stamp: 0
>>>>> [  101.733716] hardirqs last  enabled at (0): [<0000000000000000>] 0x0
>>>>> [  101.739973] hardirqs last disabled at (0): [<ffffd3ebecb10120>] copy_process+0xa78/0x1a98
>>>>> [  101.748146] softirqs last  enabled at (0): [<ffffd3ebecb10120>] copy_process+0xa78/0x1a98
>>>>> [  101.756313] softirqs last disabled at (0): [<0000000000000000>] 0x0
>>>>> [  101.762569] CPU: 4 PID: 1534 Comm: sh Not tainted 5.14.0-rc5+ #272
>>>>> [  101.774558] Call trace:
>>>>> [  101.794734]  __might_sleep+0x50/0x88
>>>>> [  101.798297]  __mutex_lock+0x60/0x938
>>>>> [  101.801863]  mutex_lock_nested+0x38/0x50
>>>>> [  101.805775]  kernfs_remove+0x2c/0x50             <---- this takes mutex_lock(&kernfs_mutex);
>>>>> [  101.809341]  sysfs_remove_dir+0x54/0x70
>>>>
>>>> The __kobject_del() occurs only if the refcount on the node
>>>> becomes zero.  This should never be true when of_find_compatible_node()
>>>> calls of_node_put() unless a "from" node is passed to of_find_compatible_node().
>>>
>>> I figured that was the assumption, that the of_node_put would never
>>> trigger a sysfs file / kobject deletion from there.
>>>
>>>> In both sja1105_mdiobus_base_tx_register() and sja1105_mdioux_base_t1_register()
>>>> a from node ("mdio") is passed to of_find_compatible_node() without first doing an
>>>> of_node_get(mdio).  If you add the of_node_get() calls the problem should be fixed.
>>>
>>> The answer seems simple enough, but stupid question, but why does
>>> of_find_compatible_node call of_node_put on "from" in the first place?
>>
>> Actually a good question.
>>
>> I do not know why of_find_compatible_node() calls of_node_put() instead of making
>> the caller of of_find_compatible_node() responsible.  That pattern was created
>> long before I was involved in devicetree and I have not gone back to read the
>> review comments of when that code was created.
> 

> Because it is an iterator function and they all drop the ref from the
> prior iteration.

That is what I was expecting before reading through the code.  But instead
I found of_find_compatible_node():

        raw_spin_lock_irqsave(&devtree_lock, flags);
        for_each_of_allnodes_from(from, np)
                if (__of_device_is_compatible(np, compatible, type, NULL) &&
                    of_node_get(np))
                        break;
        of_node_put(from);
        raw_spin_unlock_irqrestore(&devtree_lock, flags);


for_each_of_allnodes_fromir:

#define for_each_of_allnodes_from(from, dn) \
        for (dn = __of_find_all_nodes(from); dn; dn = __of_find_all_nodes(dn))


and __of_find_all_nodes() is:

struct device_node *__of_find_all_nodes(struct device_node *prev)
{
        struct device_node *np;
        if (!prev) {
                np = of_root;
        } else if (prev->child) {
                np = prev->child;
        } else {
                /* Walk back up looking for a sibling, or the end of the structure */
                np = prev;
                while (np->parent && !np->sibling)
                        np = np->parent;
                np = np->sibling; /* Might be null at the end of the tree */
        }
        return np;
}


So the iterator is not using of_node_get() and of_node_put() for each
node that is traversed.  The protection against a node disappearing
during the iteration is provided by holding devtree_lock.

> 
> I would say any open coded call where from is not NULL is an error.

I assume you mean any open coded call of of_find_compatible_node().  There are
at least a couple of instances of that.  I did only a partial grep while looking
at Vladimir's issue.

Doing the full grep now, I see 13 instances of architecture and driver code calling
of_find_compatible_node().

> It's not reliable because the DT search order is not defined and could
> change. Someone want to write a coccinelle script to check that?
> 

> The above code should be using of_get_compatible_child() instead.

Yes, of_get_compatible_child() should be used here.  Thanks for pointing
that out.

There are 13 instances of architecture and driver code calling
of_find_compatible_node().  If possible, it would be good to change all of
them to of_get_compatible_child().  If we could replace all driver
usage of of_find_compatible_node() with a from parameter of NULL to
a new wrapper without a from parameter, where the wrapper calls
of_find_compatible_node() with the from parameter set to NULL, then
we could prevent this problem from recurring.

(I did not look at all 13 instances yet, to see if this can be done.)

> 
> Rob
>